BenchHandoff v0.5.0 is a GitHub-only early release of the fail-closed,
resumable experiment-recovery CLI and its optional source-only Kubernetes
AgentRun controller.
Highlights:
- Windows recovery now uses extended-length paths for deterministic quarantine
destinations beyond the legacyMAX_PATHboundary. - The reference controller uses exactly two manager replicas and one
pre-created namespaced Lease. Its separate leader-election Role is restricted
togetandupdateon that exact Lease. - Deterministic fake-client fault injection covers create success,
AlreadyExists, committed-response loss, and status-conflict convergence. - The pinned single-node kind gate separately covers one live-
starttakeover
and one terminal-resume/pending-status takeover. It requires the same
measured Job and Pod identities, the pre-existing passive manager acquiring
exactly the next Lease transition without a restart, restored business RBAC,
and one final Job and Pod for each ofstart,resume, andverify.
Exact public evidence:
- Release commit:
c9837dc78ef199a52c23df101285c1e4d27c01e3 - Main CI: https://github.com/tiramitree/benchhandoff/actions/runs/30472827326
- Main real-kind gate:
https://github.com/tiramitree/benchhandoff/actions/runs/30472840631 - Annotated-tag CI:
https://github.com/tiramitree/benchhandoff/actions/runs/30473618474 - Annotated-tag real-kind gate:
https://github.com/tiramitree/benchhandoff/actions/runs/30473617993
The wheel, source distribution, and distribution checksum are the unchanged
bytes built and tested by the annotated-tag CI. The schema-2 takeover record
and its distinct checksum are the privacy-gated bytes from the annotated-tag
real-kind run. No controller image, Helm chart, package-registry publication,
production support, independent review, external use, or adoption is claimed.
This remains a maintainer-controlled, synthetic, fixed single-node experiment
with the API server and storage available. It does not establish strict
fencing, network-partition safety, arbitrary Pod recovery, multi-node or
multi-cluster behavior, exactly-once execution, production high availability,
performance, external adoption, or recruiting outcomes.
Development and documentation were AI-assisted. The claims above are limited
to the linked source revision, automated checks, released artifacts, and stated
boundaries.